Cache Model.matrices and trim the matrix export path #1008

Describe the feature you'd like to see

Child of #972 (phase 2, PR 5). This helps dense models too, not only sparse models.

Export now takes about 65% of the build time of a sparse model. Most of that time is in HiGHS itself, but three costs come from linopy:

1. Model.matrices builds a new MatrixAccessor on every access (linopy/model.py:361-363). Every caller that reads m.matrices pays for a full rebuild.
2. Every access runs variable_scaling_lookup and constraint_scaling_lookup, even when the model sets no scaling. This costs about 30 ms and 140 MB of temporary memory for 1.44M variables.
3. _stack (linopy/matrices.py:25-38) calls eliminate_zeros on the stacked matrix. Frozen blocks are already pruned, so this call does nothing useful (9 to 15 ms).

import numpy as np, pandas as pd, linopy
linopy.options["semantics"] = "v1"
i, t = pd.RangeIndex(2000, name="i"), pd.RangeIndex(720, name="t")
m = linopy.Model(sparse=True)
x = m.add_variables(0, 1, coords=[i, t], name="x")
m.add_constraints(x <= 1, name="ub")
m.add_objective(x.sum())
assert m.matrices is not m.matrices  # rebuilt on each access, 55 ms each here

Proposed direction

• Cache the accessor only when every constraint is frozen. This is always the case in a sparse model, and it avoids staleness after in-place edits of an unfrozen Constraint. Invalidate the cache wherever label_index.invalidate() already runs (linopy/common.py:947) and when the objective changes.
• Skip the scaling lookups when the model sets no scaling.
• Skip eliminate_zeros when all stacked blocks are CSRConstraint.
• Add the PyPSA-like benchmark from Tracking: sparse/CSR path follow-ups after #961 #972 as a tracked script under benchmark/, so this PR and the next ones can show numbers before and after.